While watching Sir Demi Hamus discuss how he created deep think, in the game theory that was behind highlights how humans let emotion get in the way of winning. The machine learning systems figured out that mobility was more important than quantity and sacrificing property is an acceptable outcome. For many humans this is an unthinkable outcome.

Because AI systems are based on neural networks and logic, they need goals and rules. We’ve built systems based on rules ever since the Turing machine and Babbages calculator. Now we are adding goals to test “intelligence”.  The surprising thing to some was that the AI sought the “mean” and ignored rules to improve the performance towards achieveing the goals.

Beyond the three main bounds of AI systems, there are additional constructive limitations due to these bounds many don’t understand until it is too late.  A human can understand senitment and therefore intent.  A maching can not (yet).  You can’t program every single thought a driver has into a machine running an autonomous car.  There is no rule about “don’t run over squirrels” for instance.  Some people do, some avoid them.  AI doesn’t know anything either way until there is a rule or goal governing that event.

The three boundaries and underlying limitations all can be used in favor of business applications. By generating better bounds and a simple rule that if a customer “goes off the reservation” with a completely unrelated question, the AI simply redirect back into the boundaries of its knowledge. If you don’t give it a rule to remain within the boundaries and bring the questions back into the boundaries the potential for hallucination exists.